I build the harness that runs AI agents unattended, and post the real results.

Pinned Tweet
the finale. after beating the game the ai went into cerulean cave and caught mewtwo, the strongest pokemon in red. master ball, one throw. that's a wrap on run one
30
9
346
66,232
the ai got its first badge. brock's team lasted two vine whips.
2
19
712
muse chose to challenge brock after training leafy and healing the team. rook handled the battle inputs and menus. leafy beat geodude and onix with one vine whip each. menu delays are cut down.
3
93
the ai trained for one move before brock. leafy finally learned vine whip.
1
14
905
muse chose to train leafy to level 13 before taking on brock. rook handled the battles and menus, and this is the battle where leafy learned it. the waiting is cut down.
1
3
113
the ai finally got out of viridian forest. leafy made it to pewter with 13 hp left.
1
16
684
muse chose to head north toward pewter. rook handled the trainer fight and the route out. this starts near the end of the battle, with some walking cut out.
5
104
the ai picked a trainer fight to get ready for brock. leafy finally hit level 6
1
18
1,014
muse chose the bug catcher for xp. rook’s code handled the moves, and leafy knocked out a level 7 weedle. the middle of the battle is cut here and the trainer fight wasn’t over yet.
3
101
the ai went looking for a bug pokemon. it caught a weedle and named it mothra
2
1
22
1,120
muse wanted a bug that would evolve quickly and picked the name before the encounter. rook’s code handled the catch and typed mothra, then checked the game’s memory to confirm the name
3
133
the ai wanted backup for the team. it got a second pidgey and named it mossy
1
1
21
842
muse asked for “backup depth” and picked the name before the encounter. it didn’t request a specific pokemon. rook’s code handled the catch and typed mossy, then checked the game’s memory to confirm the name
3
153
the ai picked a name before it even found a pokemon. it caught a pidgey and named it voltie
6
22
2,482
muse names a catch when it decides to go catching, before anything shows up. rook's code threw the balls, the first broke free and the second worked, then it typed voltie and checked the game's memory before counting it. the next pidgey got named mossy
5
240
the ai tried to buy poke balls before running oak's errand. the shop wouldn't sell until it delivered the package, and the package was a poke ball
1
14
1,098
muse wanted 5 poke balls first. rook tried the counter over and over, then sent back a plain reason: the shop won't serve you while oak's parcel is in your bag. muse's next call walked back to pallet town and delivered it. the clip skips the walk
3
157
the ai picked bulbasaur for brock. the rival picked charmander and won with 5 hp left
1
1
18
1,040
battles are rook's code, not the model. muse made zero calls during this fight
3
201
the ai got to name its starter. it went with leafy
5
1
25
1,190
the name came from muse spark in the same decision as the pick. rook's code typed it, then checked the game's memory to make sure the party really says leafy before counting it
1
6
252